Tenstorrent has entered the AI hardware market with a bold claim: the TT-QuietBox 2 is the first desktop workstation built entirely around RISC-V architecture, designed to run large-scale models without the need for cloud dependencies or server racks.
The system, which starts at $9,999, combines four Blackhole AI chips—each featuring 16 big RISC-V cores and up to 32 GB of GDDR6 memory—into a single, liquid-cooled enclosure. This configuration allows the workstation to handle models with up to 120 billion parameters locally, while also integrating 256 GB of DDR5 system memory for broader workload support.
Redefining On-Premise AI
The TT-QuietBox 2 is positioned as a tool for developers and small-to-medium businesses that require full control over their AI infrastructure. Unlike traditional GPUs, which often rely on proprietary software stacks, Tenstorrent emphasizes complete open-source visibility—from the compiler to the kernel—allowing users to debug at the hardware level and customize the system for specific needs.
- Silicon Innovation: Four Blackhole ASICs work as a unified mesh, delivering 2,654 TFLOPS at BlockFP8 precision while avoiding DRAM bottlenecks common in conventional AI hardware.
- Memory & Performance: 128 GB of GDDR6 VRAM and 256 GB of DDR5 system memory ensure high throughput without relying on HBM, which has been a supply constraint for other AI accelerators.
- Power Efficiency: The liquid-cooled design reduces idle power consumption by approximately 50% compared to previous generations, making it suitable for sustained desk-based operation.
The workstation runs Ubuntu 24.04 out of the box and supports models from PyTorch, ONNX, TensorFlow, JAX, and PaddlePaddle through TT-Forge, Tenstorrent’s open-source AI compiler. This flexibility extends to scientific research, where it can predict protein structures in under a minute—matching or exceeding the performance of high-end GPUs at a fraction of the cost.
A Shift in AI Hardware
While NVIDIA and AMD continue to dominate the GPU market with their RTX 3060 and RTX 5050 offerings, Tenstorrent’s approach targets a different segment: those who prioritize control, transparency, and cost efficiency over raw performance. The TT-QuietBox 2 avoids the memory supply shortages plaguing HBM-based systems by leveraging GDDR6 and on-chip SRAM, which could stabilize pricing in an otherwise volatile market.
For creators and researchers, this means a potential reduction in operational costs—no need for expensive server setups or cloud subscriptions. Instead, the workstation promises to deliver high-performance inference directly on a desk, with no specialized electrical requirements beyond a standard 120V outlet.
The system is expected to ship globally in Q2 2026, though availability details are not yet confirmed. If it delivers on its promises, the TT-QuietBox 2 could redefine how AI workloads are handled outside data centers—without relying on the proprietary ecosystems that currently dominate the industry.
